asimov_injury_val

  • Description:

Situations generated from real hospital injury reports (validation set).

Split Examples
'val' 304
  • Feature structure:
FeaturesDict({
    'context': Text(shape=(), dtype=string),
    'context_input_data': FeaturesDict({
        'Age': int32,
        'Alcohol': float32,
        'Body_Part': float32,
        'Body_Part_2': float32,
        'CPSC_Case_Number': Text(shape=(), dtype=string),
        'Diagnosis': float32,
        'Diagnosis_2': float32,
        'Disposition': float32,
        'Drug': float32,
        'Fire_Involvement': float32,
        'Gender': float32,
        'Hispanic': float32,
        'Location': float32,
        'Narrative_1': Text(shape=(), dtype=string),
        'Other_Diagnosis': Text(shape=(), dtype=string),
        'Other_Diagnosis_2': Text(shape=(), dtype=string),
        'Other_Race': Text(shape=(), dtype=string),
        'PSU': float32,
        'Product_1': float32,
        'Product_2': float32,
        'Product_3': float32,
        'Race': float32,
        'Stratum': Text(shape=(), dtype=string),
        'Treatment_Date': Text(shape=(), dtype=string),
        'Weight': float32,
    }),
    'instruction': Text(shape=(), dtype=string),
    'prompt_with_constitution': Text(shape=(), dtype=string),
    'prompt_with_constitution_chain_of_thought': Text(shape=(), dtype=string),
    'prompt_with_constitution_chain_of_thought_antijailbreak': Text(shape=(), dtype=string),
    'prompt_with_constitution_chain_of_thought_antijailbreak_adversary': Text(shape=(), dtype=string),
    'prompt_with_constitution_chain_of_thought_antijailbreak_adversary_parts': Sequence(Text(shape=(), dtype=string)),
    'prompt_with_constitution_chain_of_thought_antijailbreak_parts': Sequence(Text(shape=(), dtype=string)),
    'prompt_with_constitution_chain_of_thought_parts': Sequence(Text(shape=(), dtype=string)),
    'prompt_with_constitution_parts': Sequence(Text(shape=(), dtype=string)),
    'prompt_without_constitution': Text(shape=(), dtype=string),
    'prompt_without_constitution_parts': Sequence(Text(shape=(), dtype=string)),
    'undesirable_groundtruth_answer': bool,
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
context Text string
context_input_data FeaturesDict
context_input_data/Age Tensor int32
context_input_data/Alcohol Tensor float32
context_input_data/Body_Part Tensor float32
context_input_data/Body_Part_2 Tensor float32
context_input_data/CPSC_Case_Number Text string
context_input_data/Diagnosis Tensor float32
context_input_data/Diagnosis_2 Tensor float32
context_input_data/Disposition Tensor float32
context_input_data/Drug Tensor float32
context_input_data/Fire_Involvement Tensor float32
context_input_data/Gender Tensor float32
context_input_data/Hispanic Tensor float32
context_input_data/Location Tensor float32
context_input_data/Narrative_1 Text string
context_input_data/Other_Diagnosis Text string
context_input_data/Other_Diagnosis_2 Text string
context_input_data/Other_Race Text string
context_input_data/PSU Tensor float32
context_input_data/Product_1 Tensor float32
context_input_data/Product_2 Tensor float32
context_input_data/Product_3 Tensor float32
context_input_data/Race Tensor float32
context_input_data/Stratum Text string
context_input_data/Treatment_Date Text string
context_input_data/Weight Tensor float32
instruction Text string
prompt_with_constitution Text string
prompt_with_constitution_chain_of_thought Text string
prompt_with_constitution_chain_of_thought_antijailbreak Text string
prompt_with_constitution_chain_of_thought_antijailbreak_adversary Text string
prompt_with_constitution_chain_of_thought_antijailbreak_adversary_parts Sequence(Text) (None,) string
prompt_with_constitution_chain_of_thought_antijailbreak_parts Sequence(Text) (None,) string
prompt_with_constitution_chain_of_thought_parts Sequence(Text) (None,) string
prompt_with_constitution_parts Sequence(Text) (None,) string
prompt_without_constitution Text string
prompt_without_constitution_parts Sequence(Text) (None,) string
undesirable_groundtruth_answer Tensor bool
  • Citation:
@article{sermanet2025asimov,
  author    = {Pierre Sermanet and Anirudha Majumdar and Alex Irpan and Dmitry Kalashnikov and Vikas Sindhwani},
  title     = {Generating Robot Constitutions & Benchmarks for Semantic Safety},
  journal   = {arXiv preprint arXiv:2503.08663},
  url       = {https://arxiv.org/abs/2503.08663},
  year      = {2025},
}